Meta Superintelligence Lab Releases Muse Spark: A Multimodal Reasoning Model With Thought Compression and Parallel Agents

Meta Superintelligence Labs unveiled Muse Spark, the first model in Meta’s Muse family, built for native multimodal reasoning and tool use.

Meta says the system scales through pretraining, reinforcement learning, and test-time reasoning, with a rebuilt stack delivering about 10x better compute efficiency than Llama 4 Maverick.

Muse Spark adds visual chain-of-thought, parallel agent orchestration, and thought compression to solve tasks with fewer reasoning tokens.

The model showed strong results on visual localization benchmarks like ScreenSpot Pro, especially when paired with Python tools.